Sources of Variation in Cell-Type RNA-Seq Profiles

Johan Gustafsson, Felix Held, Jonathan L. Robinson, Elias Björnson, Rebecka Jörnsten, Jens Nielsen

TL;DR

This study reveals that technical variation between laboratories is often as large as the biological difference between major cell types (like B and T cells). By analyzing diverse RNA-Seq datasets, the authors identify "UMI copy fraction"—the number of duplicate reads per molecule—as a primary driver of the discrepancy between single-cell and bulk sequencing data.

What Problem Does This Paper Address?

In transcriptomics, "bulk RNA-Seq" is like analyzing a fruit smoothie: you get the average signal of everything mixed together. To understand the individual "fruits" (cell types), researchers use computational tools for deconvolution or digital cytometry. These methods rely on reference Gene Expression Profiles (GEPs)—essentially a "signature" of what a pure B cell or T cell looks like.

The problem? These signatures aren't universal. A B cell sequenced in Lab A might look more like a T cell from Lab A than a B cell from Lab B. This variation stems from a mix of biological factors (tissue of origin, cell subtype, individual genetics) and technical "batch effects" (sequencing platform, library prep, lab protocols). This paper sets out to quantify exactly which of these factors contributes most to the "noise" that complicates cross-study comparisons.

Methods

The researchers employed a multi-pronged analytical approach to dissect variation:

  1. Normalization Benchmarking: They compared standard methods like TPM (Transcripts Per Million), TMM (Trimmed Mean of M-values), and Quantile normalization to see which best handled the integration of bulk and single-cell data.
  2. LFCSD (Log Fold Change Standard Deviation): A pairwise metric used to measure the dissimilarity between two samples that differ by only one factor (e.g., same cell type, different lab).
  3. Random Forest Regression: They used Random Forests on PCA-transformed data to rank the "importance" of different covariates (lab, cell type, tissue, etc.) in explaining the total variance.
  4. Covariate Regression: To bridge the gap between UMI-based single-cell data and length-normalized bulk data, they tested regressing out factors like transcript length, GC content, and a novel metric: UMICF (UMI Copy Fraction), which represents the fraction of reads filtered out as PCR duplicates.

Datasets

The study utilized a massive collection of 105 samples (74 bulk, 31 pooled single-cell populations) from several high-profile sources:

  • Single-Cell: Human Cell Atlas (HCA), 10x Genomics (PBMC68k, B10k), and Smart-Seq2 data (MEL).
  • Bulk: ENCODE, FANTOM5, BLUEPRINT, and the EVAL dataset (which contains both bulk and 10x data from the same mouse brain samples).
  • Focus: The analysis primarily focused on B and T cells to provide a clear biological baseline for comparison.

Key Insights and Ideas

  • The "Lab Effect" is Massive: Perhaps the most striking finding is that the technical variation introduced by different laboratories is of the same magnitude as the biological variation between B cells and T cells. This suggests that using a "universal" reference profile from a different lab without correction is highly risky.
  • Individual Differences are Negligible: Compared to cell type or lab protocol, the genetic differences between individuals contributed very little to the overall variation in gene expression.
  • The UMI Duplicate Clue: The researchers found that the difference between single-cell (UMI-based) and bulk RNA-Seq is largely explained by the number of read duplicates per mRNA molecule. This "UMI copy fraction" likely captures underlying PCR amplification biases that are corrected in single-cell data but remain "baked into" bulk data.
  • Normalization Matters: TPM/CPM normalization—the industry standard for years—was found to be inadequate for comparing across technologies. TMM applied to "pseudo-counts" (TPM values scaled back to library size) performed significantly better.

Key Results

  • Variation Ranking: Lab and Cell Type were the dominant factors. Tissue of origin and cell subtype were secondary but still substantial.
  • Batch Correction: Tools like ComBat were highly effective at removing lab-to-lab differences, but the authors warn that these tools require overlapping biological samples across batches, which isn't always possible in deconvolution.
  • Correlation Improvement: By regressing out UMICF and GC content, the researchers significantly improved the correlation between single-cell and bulk expression profiles. In the EVAL dataset, Pearson correlation jumped from ~0.80 to over 0.86 after correcting for these technical covariates.

Limitations and Future Work

The authors acknowledge a few key constraints:

  • Cell Type Scope: The study focused almost exclusively on B and T cells. While these are standard models, the "technical vs. biological" ratio might look different for more similar cell types (e.g., different flavors of neurons).
  • Dataset Balance: Most samples originated from blood, which might bias the "tissue of origin" importance ranking.
  • Marker Genes: The study looked at the whole transcriptome. Many deconvolution algorithms only care about specific "marker genes," which might be more (or less) robust to technical noise than the average gene.

Future Direction: The authors suggest that creating a "library of factor patterns" from many single-cell experiments could eventually allow researchers to predict and correct for biases in bulk data more generalized way, making deconvolution far more accurate across different studies.

Summary updated

Abstract

Abstract Background Cell-type specific gene expression profiles are needed for many computational methods operating on bulk RNA-Seq samples, such as deconvolution of cell-type fractions and digital cytometry. However, the gene expression profile of a cell type can vary substantially due to both technical factors and biological differences in cell state and surroundings, reducing the efficacy of such methods. Here, we investigated which factors contribute most to this variation. Results We evaluated different normalization methods, quantified the magnitude of variation introduced by different sources, and examined the differences between UMI-based single-cell RNA-Seq and bulk RNA-Seq. We applied methods such as random forest regression to a collection of publicly available bulk and single-cell RNA-Seq datasets containing B and T cells, and found that the technical variation across laboratories is of the same magnitude as the biological variation across cell types. Tissue of origin and cell subtype are less important but still substantial factors, while the difference between individuals is relatively small. We also show that much of the differences between UMI-based single-cell and bulk RNA-Seq methods can be explained by the number of read duplicates per mRNA molecule in the single-cell sample.Conclusions Our work shows the importance of either matching or correcting for technical factors when creating cell-type specific gene expression profiles that are to be used together with bulk samples.
BibTeX
@inproceedings{gustafsson2020sources,
  author = {Johan Gustafsson and Felix Held and Jonathan L. Robinson and Elias Björnson and Rebecka Jörnsten and Jens Nielsen},
  title = {{Sources of Variation in Cell-Type RNA-Seq Profiles}},
  booktitle = {PLoS},
  year = {2020},
  doi = {10.1371/journal.pone.0239495},
}
Figure